1 /*
2  * Copyright (c) 2012-2016 VMware, Inc.  All rights reserved.
3  *
4  * This program is free software; you can redistribute it and/or
5  * modify it under the terms of EITHER the GNU General Public License
6  * version 2 as published by the Free Software Foundation or the BSD
7  * 2-Clause License. This program is distributed in the hope that it
8  * will be useful, but WITHOUT ANY WARRANTY; WITHOUT EVEN THE IMPLIED
9  * WARRANTY OF MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE.
10  * See the GNU General Public License version 2 for more details at
11  * http://www.gnu.org/licenses/old-licenses/gpl-2.0.en.html.
12  *
13  * You should have received a copy of the GNU General Public License
14  * along with this program available in the file COPYING in the main
15  * directory of this source tree.
16  *
17  * The BSD 2-Clause License
18  *
19  *     Redistribution and use in source and binary forms, with or
20  *     without modification, are permitted provided that the following
21  *     conditions are met:
22  *
23  *      - Redistributions of source code must retain the above
24  *        copyright notice, this list of conditions and the following
25  *        disclaimer.
26  *
27  *      - Redistributions in binary form must reproduce the above
28  *        copyright notice, this list of conditions and the following
29  *        disclaimer in the documentation and/or other materials
30  *        provided with the distribution.
31  *
32  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
33  * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
34  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
35  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
36  * COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT,
37  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
38  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
39  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
40  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
41  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
42  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
43  * OF THE POSSIBILITY OF SUCH DAMAGE.
44  */
45 
46 #include <asm/page.h>
47 #include <linux/io.h>
48 #include <linux/wait.h>
49 #include <rdma/ib_addr.h>
50 #include <rdma/ib_smi.h>
51 #include <rdma/ib_user_verbs.h>
52 
53 #include "pvrdma.h"
54 
55 static void __pvrdma_destroy_qp(struct pvrdma_dev *dev,
56 				struct pvrdma_qp *qp);
57 
58 static inline void get_cqs(struct pvrdma_qp *qp, struct pvrdma_cq **send_cq,
59 			   struct pvrdma_cq **recv_cq)
60 {
61 	*send_cq = to_vcq(qp->ibqp.send_cq);
62 	*recv_cq = to_vcq(qp->ibqp.recv_cq);
63 }
64 
65 static void pvrdma_lock_cqs(struct pvrdma_cq *scq, struct pvrdma_cq *rcq,
66 			    unsigned long *scq_flags,
67 			    unsigned long *rcq_flags)
68 	__acquires(scq->cq_lock) __acquires(rcq->cq_lock)
69 {
70 	if (scq == rcq) {
71 		spin_lock_irqsave(&scq->cq_lock, *scq_flags);
72 		__acquire(rcq->cq_lock);
73 	} else if (scq->cq_handle < rcq->cq_handle) {
74 		spin_lock_irqsave(&scq->cq_lock, *scq_flags);
75 		spin_lock_irqsave_nested(&rcq->cq_lock, *rcq_flags,
76 					 SINGLE_DEPTH_NESTING);
77 	} else {
78 		spin_lock_irqsave(&rcq->cq_lock, *rcq_flags);
79 		spin_lock_irqsave_nested(&scq->cq_lock, *scq_flags,
80 					 SINGLE_DEPTH_NESTING);
81 	}
82 }
83 
84 static void pvrdma_unlock_cqs(struct pvrdma_cq *scq, struct pvrdma_cq *rcq,
85 			      unsigned long *scq_flags,
86 			      unsigned long *rcq_flags)
87 	__releases(scq->cq_lock) __releases(rcq->cq_lock)
88 {
89 	if (scq == rcq) {
90 		__release(rcq->cq_lock);
91 		spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
92 	} else if (scq->cq_handle < rcq->cq_handle) {
93 		spin_unlock_irqrestore(&rcq->cq_lock, *rcq_flags);
94 		spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
95 	} else {
96 		spin_unlock_irqrestore(&scq->cq_lock, *scq_flags);
97 		spin_unlock_irqrestore(&rcq->cq_lock, *rcq_flags);
98 	}
99 }
100 
101 static void pvrdma_reset_qp(struct pvrdma_qp *qp)
102 {
103 	struct pvrdma_cq *scq, *rcq;
104 	unsigned long scq_flags, rcq_flags;
105 
106 	/* Clean up cqes */
107 	get_cqs(qp, &scq, &rcq);
108 	pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
109 
110 	_pvrdma_flush_cqe(qp, scq);
111 	if (scq != rcq)
112 		_pvrdma_flush_cqe(qp, rcq);
113 
114 	pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
115 
116 	/*
117 	 * Reset queuepair. The checks are because usermode queuepairs won't
118 	 * have kernel ringstates.
119 	 */
120 	if (qp->rq.ring) {
121 		atomic_set(&qp->rq.ring->cons_head, 0);
122 		atomic_set(&qp->rq.ring->prod_tail, 0);
123 	}
124 	if (qp->sq.ring) {
125 		atomic_set(&qp->sq.ring->cons_head, 0);
126 		atomic_set(&qp->sq.ring->prod_tail, 0);
127 	}
128 }
129 
130 static int pvrdma_set_rq_size(struct pvrdma_dev *dev,
131 			      struct ib_qp_cap *req_cap,
132 			      struct pvrdma_qp *qp)
133 {
134 	if (req_cap->max_recv_wr > dev->dsr->caps.max_qp_wr ||
135 	    req_cap->max_recv_sge > dev->dsr->caps.max_sge) {
136 		dev_warn(&dev->pdev->dev, "recv queue size invalid\n");
137 		return -EINVAL;
138 	}
139 
140 	qp->rq.wqe_cnt = roundup_pow_of_two(max(1U, req_cap->max_recv_wr));
141 	qp->rq.max_sg = roundup_pow_of_two(max(1U, req_cap->max_recv_sge));
142 
143 	/* Write back */
144 	req_cap->max_recv_wr = qp->rq.wqe_cnt;
145 	req_cap->max_recv_sge = qp->rq.max_sg;
146 
147 	qp->rq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_rq_wqe_hdr) +
148 					     sizeof(struct pvrdma_sge) *
149 					     qp->rq.max_sg);
150 	qp->npages_recv = (qp->rq.wqe_cnt * qp->rq.wqe_size + PAGE_SIZE - 1) /
151 			  PAGE_SIZE;
152 
153 	return 0;
154 }
155 
156 static int pvrdma_set_sq_size(struct pvrdma_dev *dev, struct ib_qp_cap *req_cap,
157 			      struct pvrdma_qp *qp)
158 {
159 	if (req_cap->max_send_wr > dev->dsr->caps.max_qp_wr ||
160 	    req_cap->max_send_sge > dev->dsr->caps.max_sge) {
161 		dev_warn(&dev->pdev->dev, "send queue size invalid\n");
162 		return -EINVAL;
163 	}
164 
165 	qp->sq.wqe_cnt = roundup_pow_of_two(max(1U, req_cap->max_send_wr));
166 	qp->sq.max_sg = roundup_pow_of_two(max(1U, req_cap->max_send_sge));
167 
168 	/* Write back */
169 	req_cap->max_send_wr = qp->sq.wqe_cnt;
170 	req_cap->max_send_sge = qp->sq.max_sg;
171 
172 	qp->sq.wqe_size = roundup_pow_of_two(sizeof(struct pvrdma_sq_wqe_hdr) +
173 					     sizeof(struct pvrdma_sge) *
174 					     qp->sq.max_sg);
175 	/* Note: one extra page for the header. */
176 	qp->npages_send = PVRDMA_QP_NUM_HEADER_PAGES +
177 			  (qp->sq.wqe_cnt * qp->sq.wqe_size + PAGE_SIZE - 1) /
178 								PAGE_SIZE;
179 
180 	return 0;
181 }
182 
183 /**
184  * pvrdma_create_qp - create queue pair
185  * @ibqp: queue pair
186  * @init_attr: queue pair attributes
187  * @udata: user data
188  *
189  * @return: the 0 on success, otherwise returns an errno.
190  */
191 int pvrdma_create_qp(struct ib_qp *ibqp, struct ib_qp_init_attr *init_attr,
192 		     struct ib_udata *udata)
193 {
194 	struct pvrdma_qp *qp = to_vqp(ibqp);
195 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
196 	union pvrdma_cmd_req req;
197 	union pvrdma_cmd_resp rsp;
198 	struct pvrdma_cmd_create_qp *cmd = &req.create_qp;
199 	struct pvrdma_cmd_create_qp_resp *resp = &rsp.create_qp_resp;
200 	struct pvrdma_cmd_create_qp_resp_v2 *resp_v2 = &rsp.create_qp_resp_v2;
201 	struct pvrdma_create_qp ucmd;
202 	struct pvrdma_create_qp_resp qp_resp = {};
203 	unsigned long flags;
204 	int ret;
205 	bool is_srq = !!init_attr->srq;
206 
207 	if (init_attr->create_flags) {
208 		dev_warn(&dev->pdev->dev,
209 			 "invalid create queuepair flags %#x\n",
210 			 init_attr->create_flags);
211 		return -EOPNOTSUPP;
212 	}
213 
214 	if (init_attr->qp_type != IB_QPT_RC &&
215 	    init_attr->qp_type != IB_QPT_UD &&
216 	    init_attr->qp_type != IB_QPT_GSI) {
217 		dev_warn(&dev->pdev->dev, "queuepair type %d not supported\n",
218 			 init_attr->qp_type);
219 		return -EOPNOTSUPP;
220 	}
221 
222 	if (is_srq && !dev->dsr->caps.max_srq) {
223 		dev_warn(&dev->pdev->dev,
224 			 "SRQs not supported by device\n");
225 		return -EINVAL;
226 	}
227 
228 	if (!atomic_add_unless(&dev->num_qps, 1, dev->dsr->caps.max_qp))
229 		return -ENOMEM;
230 
231 	switch (init_attr->qp_type) {
232 	case IB_QPT_GSI:
233 		if (init_attr->port_num == 0 ||
234 		    init_attr->port_num > ibqp->device->phys_port_cnt) {
235 			dev_warn(&dev->pdev->dev, "invalid queuepair attrs\n");
236 			ret = -EINVAL;
237 			goto err_qp;
238 		}
239 		fallthrough;
240 	case IB_QPT_RC:
241 	case IB_QPT_UD:
242 		spin_lock_init(&qp->sq.lock);
243 		spin_lock_init(&qp->rq.lock);
244 		mutex_init(&qp->mutex);
245 		refcount_set(&qp->refcnt, 1);
246 		init_completion(&qp->free);
247 
248 		qp->state = IB_QPS_RESET;
249 		qp->is_kernel = !udata;
250 
251 		if (!qp->is_kernel) {
252 			dev_dbg(&dev->pdev->dev,
253 				"create queuepair from user space\n");
254 
255 			if (ib_copy_from_udata(&ucmd, udata, sizeof(ucmd))) {
256 				ret = -EFAULT;
257 				goto err_qp;
258 			}
259 
260 			/* Userspace supports qpn and qp handles? */
261 			if (dev->dsr_version >= PVRDMA_QPHANDLE_VERSION &&
262 			    udata->outlen < sizeof(qp_resp)) {
263 				dev_warn(&dev->pdev->dev,
264 					 "create queuepair not supported\n");
265 				ret = -EOPNOTSUPP;
266 				goto err_qp;
267 			}
268 
269 			if (!is_srq) {
270 				/* set qp->sq.wqe_cnt, shift, buf_size.. */
271 				qp->rumem = ib_umem_get(ibqp->device,
272 							ucmd.rbuf_addr,
273 							ucmd.rbuf_size, 0);
274 				if (IS_ERR(qp->rumem)) {
275 					ret = PTR_ERR(qp->rumem);
276 					goto err_qp;
277 				}
278 				qp->srq = NULL;
279 			} else {
280 				qp->rumem = NULL;
281 				qp->srq = to_vsrq(init_attr->srq);
282 			}
283 
284 			qp->sumem = ib_umem_get(ibqp->device, ucmd.sbuf_addr,
285 						ucmd.sbuf_size, 0);
286 			if (IS_ERR(qp->sumem)) {
287 				if (!is_srq)
288 					ib_umem_release(qp->rumem);
289 				ret = PTR_ERR(qp->sumem);
290 				goto err_qp;
291 			}
292 
293 			qp->npages_send =
294 				ib_umem_num_dma_blocks(qp->sumem, PAGE_SIZE);
295 			if (!is_srq)
296 				qp->npages_recv = ib_umem_num_dma_blocks(
297 					qp->rumem, PAGE_SIZE);
298 			else
299 				qp->npages_recv = 0;
300 			qp->npages = qp->npages_send + qp->npages_recv;
301 		} else {
302 			ret = pvrdma_set_sq_size(to_vdev(ibqp->device),
303 						 &init_attr->cap, qp);
304 			if (ret)
305 				goto err_qp;
306 
307 			ret = pvrdma_set_rq_size(to_vdev(ibqp->device),
308 						 &init_attr->cap, qp);
309 			if (ret)
310 				goto err_qp;
311 
312 			qp->npages = qp->npages_send + qp->npages_recv;
313 
314 			/* Skip header page. */
315 			qp->sq.offset = PVRDMA_QP_NUM_HEADER_PAGES * PAGE_SIZE;
316 
317 			/* Recv queue pages are after send pages. */
318 			qp->rq.offset = qp->npages_send * PAGE_SIZE;
319 		}
320 
321 		if (qp->npages < 0 || qp->npages > PVRDMA_PAGE_DIR_MAX_PAGES) {
322 			dev_warn(&dev->pdev->dev,
323 				 "overflow pages in queuepair\n");
324 			ret = -EINVAL;
325 			goto err_umem;
326 		}
327 
328 		ret = pvrdma_page_dir_init(dev, &qp->pdir, qp->npages,
329 					   qp->is_kernel);
330 		if (ret) {
331 			dev_warn(&dev->pdev->dev,
332 				 "could not allocate page directory\n");
333 			goto err_umem;
334 		}
335 
336 		if (!qp->is_kernel) {
337 			pvrdma_page_dir_insert_umem(&qp->pdir, qp->sumem, 0);
338 			if (!is_srq)
339 				pvrdma_page_dir_insert_umem(&qp->pdir,
340 							    qp->rumem,
341 							    qp->npages_send);
342 		} else {
343 			/* Ring state is always the first page. */
344 			qp->sq.ring = qp->pdir.pages[0];
345 			qp->rq.ring = is_srq ? NULL : &qp->sq.ring[1];
346 		}
347 		break;
348 	default:
349 		ret = -EINVAL;
350 		goto err_qp;
351 	}
352 
353 	/* Not supported */
354 	init_attr->cap.max_inline_data = 0;
355 
356 	memset(cmd, 0, sizeof(*cmd));
357 	cmd->hdr.cmd = PVRDMA_CMD_CREATE_QP;
358 	cmd->pd_handle = to_vpd(ibqp->pd)->pd_handle;
359 	cmd->send_cq_handle = to_vcq(init_attr->send_cq)->cq_handle;
360 	cmd->recv_cq_handle = to_vcq(init_attr->recv_cq)->cq_handle;
361 	if (is_srq)
362 		cmd->srq_handle = to_vsrq(init_attr->srq)->srq_handle;
363 	else
364 		cmd->srq_handle = 0;
365 	cmd->max_send_wr = init_attr->cap.max_send_wr;
366 	cmd->max_recv_wr = init_attr->cap.max_recv_wr;
367 	cmd->max_send_sge = init_attr->cap.max_send_sge;
368 	cmd->max_recv_sge = init_attr->cap.max_recv_sge;
369 	cmd->max_inline_data = init_attr->cap.max_inline_data;
370 	cmd->sq_sig_all = (init_attr->sq_sig_type == IB_SIGNAL_ALL_WR) ? 1 : 0;
371 	cmd->qp_type = ib_qp_type_to_pvrdma(init_attr->qp_type);
372 	cmd->is_srq = is_srq;
373 	cmd->lkey = 0;
374 	cmd->access_flags = IB_ACCESS_LOCAL_WRITE;
375 	cmd->total_chunks = qp->npages;
376 	cmd->send_chunks = qp->npages_send - PVRDMA_QP_NUM_HEADER_PAGES;
377 	cmd->pdir_dma = qp->pdir.dir_dma;
378 
379 	dev_dbg(&dev->pdev->dev, "create queuepair with %d, %d, %d, %d\n",
380 		cmd->max_send_wr, cmd->max_recv_wr, cmd->max_send_sge,
381 		cmd->max_recv_sge);
382 
383 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_CREATE_QP_RESP);
384 	if (ret < 0) {
385 		dev_warn(&dev->pdev->dev,
386 			 "could not create queuepair, error: %d\n", ret);
387 		goto err_pdir;
388 	}
389 
390 	/* max_send_wr/_recv_wr/_send_sge/_recv_sge/_inline_data */
391 	qp->port = init_attr->port_num;
392 
393 	if (dev->dsr_version >= PVRDMA_QPHANDLE_VERSION) {
394 		qp->ibqp.qp_num = resp_v2->qpn;
395 		qp->qp_handle = resp_v2->qp_handle;
396 	} else {
397 		qp->ibqp.qp_num = resp->qpn;
398 		qp->qp_handle = resp->qpn;
399 	}
400 
401 	spin_lock_irqsave(&dev->qp_tbl_lock, flags);
402 	dev->qp_tbl[qp->qp_handle % dev->dsr->caps.max_qp] = qp;
403 	spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
404 
405 	if (udata) {
406 		qp_resp.qpn = qp->ibqp.qp_num;
407 		qp_resp.qp_handle = qp->qp_handle;
408 
409 		if (ib_copy_to_udata(udata, &qp_resp,
410 				     min(udata->outlen, sizeof(qp_resp)))) {
411 			dev_warn(&dev->pdev->dev,
412 				 "failed to copy back udata\n");
413 			__pvrdma_destroy_qp(dev, qp);
414 			return -EINVAL;
415 		}
416 	}
417 
418 	return 0;
419 
420 err_pdir:
421 	pvrdma_page_dir_cleanup(dev, &qp->pdir);
422 err_umem:
423 	ib_umem_release(qp->rumem);
424 	ib_umem_release(qp->sumem);
425 err_qp:
426 	atomic_dec(&dev->num_qps);
427 	return ret;
428 }
429 
430 static void _pvrdma_free_qp(struct pvrdma_qp *qp)
431 {
432 	unsigned long flags;
433 	struct pvrdma_dev *dev = to_vdev(qp->ibqp.device);
434 
435 	spin_lock_irqsave(&dev->qp_tbl_lock, flags);
436 	dev->qp_tbl[qp->qp_handle] = NULL;
437 	spin_unlock_irqrestore(&dev->qp_tbl_lock, flags);
438 
439 	if (refcount_dec_and_test(&qp->refcnt))
440 		complete(&qp->free);
441 	wait_for_completion(&qp->free);
442 
443 	ib_umem_release(qp->rumem);
444 	ib_umem_release(qp->sumem);
445 
446 	pvrdma_page_dir_cleanup(dev, &qp->pdir);
447 
448 	atomic_dec(&dev->num_qps);
449 }
450 
451 static void pvrdma_free_qp(struct pvrdma_qp *qp)
452 {
453 	struct pvrdma_cq *scq;
454 	struct pvrdma_cq *rcq;
455 	unsigned long scq_flags, rcq_flags;
456 
457 	/* In case cq is polling */
458 	get_cqs(qp, &scq, &rcq);
459 	pvrdma_lock_cqs(scq, rcq, &scq_flags, &rcq_flags);
460 
461 	_pvrdma_flush_cqe(qp, scq);
462 	if (scq != rcq)
463 		_pvrdma_flush_cqe(qp, rcq);
464 
465 	/*
466 	 * We're now unlocking the CQs before clearing out the qp handle this
467 	 * should still be safe. We have destroyed the backend QP and flushed
468 	 * the CQEs so there should be no other completions for this QP.
469 	 */
470 	pvrdma_unlock_cqs(scq, rcq, &scq_flags, &rcq_flags);
471 
472 	_pvrdma_free_qp(qp);
473 }
474 
475 static inline void _pvrdma_destroy_qp_work(struct pvrdma_dev *dev,
476 					   u32 qp_handle)
477 {
478 	union pvrdma_cmd_req req;
479 	struct pvrdma_cmd_destroy_qp *cmd = &req.destroy_qp;
480 	int ret;
481 
482 	memset(cmd, 0, sizeof(*cmd));
483 	cmd->hdr.cmd = PVRDMA_CMD_DESTROY_QP;
484 	cmd->qp_handle = qp_handle;
485 
486 	ret = pvrdma_cmd_post(dev, &req, NULL, 0);
487 	if (ret < 0)
488 		dev_warn(&dev->pdev->dev,
489 			 "destroy queuepair failed, error: %d\n", ret);
490 }
491 
492 /**
493  * pvrdma_destroy_qp - destroy a queue pair
494  * @qp: the queue pair to destroy
495  * @udata: user data or null for kernel object
496  *
497  * @return: always 0.
498  */
499 int pvrdma_destroy_qp(struct ib_qp *qp, struct ib_udata *udata)
500 {
501 	struct pvrdma_qp *vqp = to_vqp(qp);
502 
503 	_pvrdma_destroy_qp_work(to_vdev(qp->device), vqp->qp_handle);
504 	pvrdma_free_qp(vqp);
505 
506 	return 0;
507 }
508 
509 static void __pvrdma_destroy_qp(struct pvrdma_dev *dev,
510 				struct pvrdma_qp *qp)
511 {
512 	_pvrdma_destroy_qp_work(dev, qp->qp_handle);
513 	_pvrdma_free_qp(qp);
514 }
515 
516 /**
517  * pvrdma_modify_qp - modify queue pair attributes
518  * @ibqp: the queue pair
519  * @attr: the new queue pair's attributes
520  * @attr_mask: attributes mask
521  * @udata: user data
522  *
523  * @returns 0 on success, otherwise returns an errno.
524  */
525 int pvrdma_modify_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
526 		     int attr_mask, struct ib_udata *udata)
527 {
528 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
529 	struct pvrdma_qp *qp = to_vqp(ibqp);
530 	union pvrdma_cmd_req req;
531 	union pvrdma_cmd_resp rsp;
532 	struct pvrdma_cmd_modify_qp *cmd = &req.modify_qp;
533 	enum ib_qp_state cur_state, next_state;
534 	int ret;
535 
536 	if (attr_mask & ~IB_QP_ATTR_STANDARD_BITS)
537 		return -EOPNOTSUPP;
538 
539 	/* Sanity checking. Should need lock here */
540 	mutex_lock(&qp->mutex);
541 	cur_state = (attr_mask & IB_QP_CUR_STATE) ? attr->cur_qp_state :
542 		qp->state;
543 	next_state = (attr_mask & IB_QP_STATE) ? attr->qp_state : cur_state;
544 
545 	if (!ib_modify_qp_is_ok(cur_state, next_state, ibqp->qp_type,
546 				attr_mask)) {
547 		ret = -EINVAL;
548 		goto out;
549 	}
550 
551 	if (attr_mask & IB_QP_PORT) {
552 		if (attr->port_num == 0 ||
553 		    attr->port_num > ibqp->device->phys_port_cnt) {
554 			ret = -EINVAL;
555 			goto out;
556 		}
557 	}
558 
559 	if (attr_mask & IB_QP_MIN_RNR_TIMER) {
560 		if (attr->min_rnr_timer > 31) {
561 			ret = -EINVAL;
562 			goto out;
563 		}
564 	}
565 
566 	if (attr_mask & IB_QP_PKEY_INDEX) {
567 		if (attr->pkey_index >= dev->dsr->caps.max_pkeys) {
568 			ret = -EINVAL;
569 			goto out;
570 		}
571 	}
572 
573 	if (attr_mask & IB_QP_QKEY)
574 		qp->qkey = attr->qkey;
575 
576 	if (cur_state == next_state && cur_state == IB_QPS_RESET) {
577 		ret = 0;
578 		goto out;
579 	}
580 
581 	qp->state = next_state;
582 	memset(cmd, 0, sizeof(*cmd));
583 	cmd->hdr.cmd = PVRDMA_CMD_MODIFY_QP;
584 	cmd->qp_handle = qp->qp_handle;
585 	cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
586 	cmd->attrs.qp_state = ib_qp_state_to_pvrdma(attr->qp_state);
587 	cmd->attrs.cur_qp_state =
588 		ib_qp_state_to_pvrdma(attr->cur_qp_state);
589 	cmd->attrs.path_mtu = ib_mtu_to_pvrdma(attr->path_mtu);
590 	cmd->attrs.path_mig_state =
591 		ib_mig_state_to_pvrdma(attr->path_mig_state);
592 	cmd->attrs.qkey = attr->qkey;
593 	cmd->attrs.rq_psn = attr->rq_psn;
594 	cmd->attrs.sq_psn = attr->sq_psn;
595 	cmd->attrs.dest_qp_num = attr->dest_qp_num;
596 	cmd->attrs.qp_access_flags =
597 		ib_access_flags_to_pvrdma(attr->qp_access_flags);
598 	cmd->attrs.pkey_index = attr->pkey_index;
599 	cmd->attrs.alt_pkey_index = attr->alt_pkey_index;
600 	cmd->attrs.en_sqd_async_notify = attr->en_sqd_async_notify;
601 	cmd->attrs.sq_draining = attr->sq_draining;
602 	cmd->attrs.max_rd_atomic = attr->max_rd_atomic;
603 	cmd->attrs.max_dest_rd_atomic = attr->max_dest_rd_atomic;
604 	cmd->attrs.min_rnr_timer = attr->min_rnr_timer;
605 	cmd->attrs.port_num = attr->port_num;
606 	cmd->attrs.timeout = attr->timeout;
607 	cmd->attrs.retry_cnt = attr->retry_cnt;
608 	cmd->attrs.rnr_retry = attr->rnr_retry;
609 	cmd->attrs.alt_port_num = attr->alt_port_num;
610 	cmd->attrs.alt_timeout = attr->alt_timeout;
611 	ib_qp_cap_to_pvrdma(&cmd->attrs.cap, &attr->cap);
612 	rdma_ah_attr_to_pvrdma(&cmd->attrs.ah_attr, &attr->ah_attr);
613 	rdma_ah_attr_to_pvrdma(&cmd->attrs.alt_ah_attr, &attr->alt_ah_attr);
614 
615 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_MODIFY_QP_RESP);
616 	if (ret < 0) {
617 		dev_warn(&dev->pdev->dev,
618 			 "could not modify queuepair, error: %d\n", ret);
619 	} else if (rsp.hdr.err > 0) {
620 		dev_warn(&dev->pdev->dev,
621 			 "cannot modify queuepair, error: %d\n", rsp.hdr.err);
622 		ret = -EINVAL;
623 	}
624 
625 	if (ret == 0 && next_state == IB_QPS_RESET)
626 		pvrdma_reset_qp(qp);
627 
628 out:
629 	mutex_unlock(&qp->mutex);
630 
631 	return ret;
632 }
633 
634 static inline void *get_sq_wqe(struct pvrdma_qp *qp, unsigned int n)
635 {
636 	return pvrdma_page_dir_get_ptr(&qp->pdir,
637 				       qp->sq.offset + n * qp->sq.wqe_size);
638 }
639 
640 static inline void *get_rq_wqe(struct pvrdma_qp *qp, unsigned int n)
641 {
642 	return pvrdma_page_dir_get_ptr(&qp->pdir,
643 				       qp->rq.offset + n * qp->rq.wqe_size);
644 }
645 
646 static int set_reg_seg(struct pvrdma_sq_wqe_hdr *wqe_hdr,
647 		       const struct ib_reg_wr *wr)
648 {
649 	struct pvrdma_user_mr *mr = to_vmr(wr->mr);
650 
651 	wqe_hdr->wr.fast_reg.iova_start = mr->ibmr.iova;
652 	wqe_hdr->wr.fast_reg.pl_pdir_dma = mr->pdir.dir_dma;
653 	wqe_hdr->wr.fast_reg.page_shift = mr->page_shift;
654 	wqe_hdr->wr.fast_reg.page_list_len = mr->npages;
655 	wqe_hdr->wr.fast_reg.length = mr->ibmr.length;
656 	wqe_hdr->wr.fast_reg.access_flags = wr->access;
657 	wqe_hdr->wr.fast_reg.rkey = wr->key;
658 
659 	return pvrdma_page_dir_insert_page_list(&mr->pdir, mr->pages,
660 						mr->npages);
661 }
662 
663 /**
664  * pvrdma_post_send - post send work request entries on a QP
665  * @ibqp: the QP
666  * @wr: work request list to post
667  * @bad_wr: the first bad WR returned
668  *
669  * @return: 0 on success, otherwise errno returned.
670  */
671 int pvrdma_post_send(struct ib_qp *ibqp, const struct ib_send_wr *wr,
672 		     const struct ib_send_wr **bad_wr)
673 {
674 	struct pvrdma_qp *qp = to_vqp(ibqp);
675 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
676 	unsigned long flags;
677 	struct pvrdma_sq_wqe_hdr *wqe_hdr;
678 	struct pvrdma_sge *sge;
679 	int i, ret;
680 
681 	/*
682 	 * In states lower than RTS, we can fail immediately. In other states,
683 	 * just post and let the device figure it out.
684 	 */
685 	if (qp->state < IB_QPS_RTS) {
686 		*bad_wr = wr;
687 		return -EINVAL;
688 	}
689 
690 	spin_lock_irqsave(&qp->sq.lock, flags);
691 
692 	while (wr) {
693 		unsigned int tail = 0;
694 
695 		if (unlikely(!pvrdma_idx_ring_has_space(
696 				qp->sq.ring, qp->sq.wqe_cnt, &tail))) {
697 			dev_warn_ratelimited(&dev->pdev->dev,
698 					     "send queue is full\n");
699 			*bad_wr = wr;
700 			ret = -ENOMEM;
701 			goto out;
702 		}
703 
704 		if (unlikely(wr->num_sge > qp->sq.max_sg || wr->num_sge < 0)) {
705 			dev_warn_ratelimited(&dev->pdev->dev,
706 					     "send SGE overflow\n");
707 			*bad_wr = wr;
708 			ret = -EINVAL;
709 			goto out;
710 		}
711 
712 		/*
713 		 * Only support UD, RC.
714 		 * Need to check opcode table for thorough checking.
715 		 * opcode		_UD	_UC	_RC
716 		 * _SEND		x	x	x
717 		 * _SEND_WITH_IMM	x	x	x
718 		 * _RDMA_WRITE			x	x
719 		 * _RDMA_WRITE_WITH_IMM		x	x
720 		 * _LOCAL_INV			x	x
721 		 * _SEND_WITH_INV		x	x
722 		 * _RDMA_READ				x
723 		 * _ATOMIC_CMP_AND_SWP			x
724 		 * _ATOMIC_FETCH_AND_ADD		x
725 		 * _MASK_ATOMIC_CMP_AND_SWP		x
726 		 * _MASK_ATOMIC_FETCH_AND_ADD		x
727 		 * _REG_MR				x
728 		 *
729 		 */
730 		if (qp->ibqp.qp_type != IB_QPT_UD &&
731 		    qp->ibqp.qp_type != IB_QPT_RC &&
732 			wr->opcode != IB_WR_SEND) {
733 			dev_warn_ratelimited(&dev->pdev->dev,
734 					     "unsupported queuepair type\n");
735 			*bad_wr = wr;
736 			ret = -EINVAL;
737 			goto out;
738 		} else if (qp->ibqp.qp_type == IB_QPT_UD ||
739 			   qp->ibqp.qp_type == IB_QPT_GSI) {
740 			if (wr->opcode != IB_WR_SEND &&
741 			    wr->opcode != IB_WR_SEND_WITH_IMM) {
742 				dev_warn_ratelimited(&dev->pdev->dev,
743 						     "invalid send opcode\n");
744 				*bad_wr = wr;
745 				ret = -EINVAL;
746 				goto out;
747 			}
748 		}
749 
750 		wqe_hdr = (struct pvrdma_sq_wqe_hdr *)get_sq_wqe(qp, tail);
751 		memset(wqe_hdr, 0, sizeof(*wqe_hdr));
752 		wqe_hdr->wr_id = wr->wr_id;
753 		wqe_hdr->num_sge = wr->num_sge;
754 		wqe_hdr->opcode = ib_wr_opcode_to_pvrdma(wr->opcode);
755 		wqe_hdr->send_flags = ib_send_flags_to_pvrdma(wr->send_flags);
756 		if (wr->opcode == IB_WR_SEND_WITH_IMM ||
757 		    wr->opcode == IB_WR_RDMA_WRITE_WITH_IMM)
758 			wqe_hdr->ex.imm_data = wr->ex.imm_data;
759 
760 		if (unlikely(wqe_hdr->opcode == PVRDMA_WR_ERROR)) {
761 			*bad_wr = wr;
762 			ret = -EINVAL;
763 			goto out;
764 		}
765 
766 		switch (qp->ibqp.qp_type) {
767 		case IB_QPT_GSI:
768 		case IB_QPT_UD:
769 			if (unlikely(!ud_wr(wr)->ah)) {
770 				dev_warn_ratelimited(&dev->pdev->dev,
771 						     "invalid address handle\n");
772 				*bad_wr = wr;
773 				ret = -EINVAL;
774 				goto out;
775 			}
776 
777 			/*
778 			 * Use qkey from qp context if high order bit set,
779 			 * otherwise from work request.
780 			 */
781 			wqe_hdr->wr.ud.remote_qpn = ud_wr(wr)->remote_qpn;
782 			wqe_hdr->wr.ud.remote_qkey =
783 				ud_wr(wr)->remote_qkey & 0x80000000 ?
784 				qp->qkey : ud_wr(wr)->remote_qkey;
785 			wqe_hdr->wr.ud.av = to_vah(ud_wr(wr)->ah)->av;
786 
787 			break;
788 		case IB_QPT_RC:
789 			switch (wr->opcode) {
790 			case IB_WR_RDMA_READ:
791 			case IB_WR_RDMA_WRITE:
792 			case IB_WR_RDMA_WRITE_WITH_IMM:
793 				wqe_hdr->wr.rdma.remote_addr =
794 					rdma_wr(wr)->remote_addr;
795 				wqe_hdr->wr.rdma.rkey = rdma_wr(wr)->rkey;
796 				break;
797 			case IB_WR_LOCAL_INV:
798 			case IB_WR_SEND_WITH_INV:
799 				wqe_hdr->ex.invalidate_rkey =
800 					wr->ex.invalidate_rkey;
801 				break;
802 			case IB_WR_ATOMIC_CMP_AND_SWP:
803 			case IB_WR_ATOMIC_FETCH_AND_ADD:
804 				wqe_hdr->wr.atomic.remote_addr =
805 					atomic_wr(wr)->remote_addr;
806 				wqe_hdr->wr.atomic.rkey = atomic_wr(wr)->rkey;
807 				wqe_hdr->wr.atomic.compare_add =
808 					atomic_wr(wr)->compare_add;
809 				if (wr->opcode == IB_WR_ATOMIC_CMP_AND_SWP)
810 					wqe_hdr->wr.atomic.swap =
811 						atomic_wr(wr)->swap;
812 				break;
813 			case IB_WR_REG_MR:
814 				ret = set_reg_seg(wqe_hdr, reg_wr(wr));
815 				if (ret < 0) {
816 					dev_warn_ratelimited(&dev->pdev->dev,
817 							     "Failed to set fast register work request\n");
818 					*bad_wr = wr;
819 					goto out;
820 				}
821 				break;
822 			default:
823 				break;
824 			}
825 
826 			break;
827 		default:
828 			dev_warn_ratelimited(&dev->pdev->dev,
829 					     "invalid queuepair type\n");
830 			ret = -EINVAL;
831 			*bad_wr = wr;
832 			goto out;
833 		}
834 
835 		sge = (struct pvrdma_sge *)(wqe_hdr + 1);
836 		for (i = 0; i < wr->num_sge; i++) {
837 			/* Need to check wqe_size 0 or max size */
838 			sge->addr = wr->sg_list[i].addr;
839 			sge->length = wr->sg_list[i].length;
840 			sge->lkey = wr->sg_list[i].lkey;
841 			sge++;
842 		}
843 
844 		/* Make sure wqe is written before index update */
845 		smp_wmb();
846 
847 		/* Update shared sq ring */
848 		pvrdma_idx_ring_inc(&qp->sq.ring->prod_tail,
849 				    qp->sq.wqe_cnt);
850 
851 		wr = wr->next;
852 	}
853 
854 	ret = 0;
855 
856 out:
857 	spin_unlock_irqrestore(&qp->sq.lock, flags);
858 
859 	if (!ret)
860 		pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_SEND | qp->qp_handle);
861 
862 	return ret;
863 }
864 
865 /**
866  * pvrdma_post_recv - post receive work request entries on a QP
867  * @ibqp: the QP
868  * @wr: the work request list to post
869  * @bad_wr: the first bad WR returned
870  *
871  * @return: 0 on success, otherwise errno returned.
872  */
873 int pvrdma_post_recv(struct ib_qp *ibqp, const struct ib_recv_wr *wr,
874 		     const struct ib_recv_wr **bad_wr)
875 {
876 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
877 	unsigned long flags;
878 	struct pvrdma_qp *qp = to_vqp(ibqp);
879 	struct pvrdma_rq_wqe_hdr *wqe_hdr;
880 	struct pvrdma_sge *sge;
881 	int ret = 0;
882 	int i;
883 
884 	/*
885 	 * In the RESET state, we can fail immediately. For other states,
886 	 * just post and let the device figure it out.
887 	 */
888 	if (qp->state == IB_QPS_RESET) {
889 		*bad_wr = wr;
890 		return -EINVAL;
891 	}
892 
893 	if (qp->srq) {
894 		dev_warn(&dev->pdev->dev, "QP associated with SRQ\n");
895 		*bad_wr = wr;
896 		return -EINVAL;
897 	}
898 
899 	spin_lock_irqsave(&qp->rq.lock, flags);
900 
901 	while (wr) {
902 		unsigned int tail = 0;
903 
904 		if (unlikely(wr->num_sge > qp->rq.max_sg ||
905 			     wr->num_sge < 0)) {
906 			ret = -EINVAL;
907 			*bad_wr = wr;
908 			dev_warn_ratelimited(&dev->pdev->dev,
909 					     "recv SGE overflow\n");
910 			goto out;
911 		}
912 
913 		if (unlikely(!pvrdma_idx_ring_has_space(
914 				qp->rq.ring, qp->rq.wqe_cnt, &tail))) {
915 			ret = -ENOMEM;
916 			*bad_wr = wr;
917 			dev_warn_ratelimited(&dev->pdev->dev,
918 					     "recv queue full\n");
919 			goto out;
920 		}
921 
922 		wqe_hdr = (struct pvrdma_rq_wqe_hdr *)get_rq_wqe(qp, tail);
923 		wqe_hdr->wr_id = wr->wr_id;
924 		wqe_hdr->num_sge = wr->num_sge;
925 		wqe_hdr->total_len = 0;
926 
927 		sge = (struct pvrdma_sge *)(wqe_hdr + 1);
928 		for (i = 0; i < wr->num_sge; i++) {
929 			sge->addr = wr->sg_list[i].addr;
930 			sge->length = wr->sg_list[i].length;
931 			sge->lkey = wr->sg_list[i].lkey;
932 			sge++;
933 		}
934 
935 		/* Make sure wqe is written before index update */
936 		smp_wmb();
937 
938 		/* Update shared rq ring */
939 		pvrdma_idx_ring_inc(&qp->rq.ring->prod_tail,
940 				    qp->rq.wqe_cnt);
941 
942 		wr = wr->next;
943 	}
944 
945 	spin_unlock_irqrestore(&qp->rq.lock, flags);
946 
947 	pvrdma_write_uar_qp(dev, PVRDMA_UAR_QP_RECV | qp->qp_handle);
948 
949 	return ret;
950 
951 out:
952 	spin_unlock_irqrestore(&qp->rq.lock, flags);
953 
954 	return ret;
955 }
956 
957 /**
958  * pvrdma_query_qp - query a queue pair's attributes
959  * @ibqp: the queue pair to query
960  * @attr: the queue pair's attributes
961  * @attr_mask: attributes mask
962  * @init_attr: initial queue pair attributes
963  *
964  * @returns 0 on success, otherwise returns an errno.
965  */
966 int pvrdma_query_qp(struct ib_qp *ibqp, struct ib_qp_attr *attr,
967 		    int attr_mask, struct ib_qp_init_attr *init_attr)
968 {
969 	struct pvrdma_dev *dev = to_vdev(ibqp->device);
970 	struct pvrdma_qp *qp = to_vqp(ibqp);
971 	union pvrdma_cmd_req req;
972 	union pvrdma_cmd_resp rsp;
973 	struct pvrdma_cmd_query_qp *cmd = &req.query_qp;
974 	struct pvrdma_cmd_query_qp_resp *resp = &rsp.query_qp_resp;
975 	int ret = 0;
976 
977 	mutex_lock(&qp->mutex);
978 
979 	if (qp->state == IB_QPS_RESET) {
980 		attr->qp_state = IB_QPS_RESET;
981 		goto out;
982 	}
983 
984 	memset(cmd, 0, sizeof(*cmd));
985 	cmd->hdr.cmd = PVRDMA_CMD_QUERY_QP;
986 	cmd->qp_handle = qp->qp_handle;
987 	cmd->attr_mask = ib_qp_attr_mask_to_pvrdma(attr_mask);
988 
989 	ret = pvrdma_cmd_post(dev, &req, &rsp, PVRDMA_CMD_QUERY_QP_RESP);
990 	if (ret < 0) {
991 		dev_warn(&dev->pdev->dev,
992 			 "could not query queuepair, error: %d\n", ret);
993 		goto out;
994 	}
995 
996 	attr->qp_state = pvrdma_qp_state_to_ib(resp->attrs.qp_state);
997 	attr->cur_qp_state =
998 		pvrdma_qp_state_to_ib(resp->attrs.cur_qp_state);
999 	attr->path_mtu = pvrdma_mtu_to_ib(resp->attrs.path_mtu);
1000 	attr->path_mig_state =
1001 		pvrdma_mig_state_to_ib(resp->attrs.path_mig_state);
1002 	attr->qkey = resp->attrs.qkey;
1003 	attr->rq_psn = resp->attrs.rq_psn;
1004 	attr->sq_psn = resp->attrs.sq_psn;
1005 	attr->dest_qp_num = resp->attrs.dest_qp_num;
1006 	attr->qp_access_flags =
1007 		pvrdma_access_flags_to_ib(resp->attrs.qp_access_flags);
1008 	attr->pkey_index = resp->attrs.pkey_index;
1009 	attr->alt_pkey_index = resp->attrs.alt_pkey_index;
1010 	attr->en_sqd_async_notify = resp->attrs.en_sqd_async_notify;
1011 	attr->sq_draining = resp->attrs.sq_draining;
1012 	attr->max_rd_atomic = resp->attrs.max_rd_atomic;
1013 	attr->max_dest_rd_atomic = resp->attrs.max_dest_rd_atomic;
1014 	attr->min_rnr_timer = resp->attrs.min_rnr_timer;
1015 	attr->port_num = resp->attrs.port_num;
1016 	attr->timeout = resp->attrs.timeout;
1017 	attr->retry_cnt = resp->attrs.retry_cnt;
1018 	attr->rnr_retry = resp->attrs.rnr_retry;
1019 	attr->alt_port_num = resp->attrs.alt_port_num;
1020 	attr->alt_timeout = resp->attrs.alt_timeout;
1021 	pvrdma_qp_cap_to_ib(&attr->cap, &resp->attrs.cap);
1022 	pvrdma_ah_attr_to_rdma(&attr->ah_attr, &resp->attrs.ah_attr);
1023 	pvrdma_ah_attr_to_rdma(&attr->alt_ah_attr, &resp->attrs.alt_ah_attr);
1024 
1025 	qp->state = attr->qp_state;
1026 
1027 	ret = 0;
1028 
1029 out:
1030 	attr->cur_qp_state = attr->qp_state;
1031 
1032 	init_attr->event_handler = qp->ibqp.event_handler;
1033 	init_attr->qp_context = qp->ibqp.qp_context;
1034 	init_attr->send_cq = qp->ibqp.send_cq;
1035 	init_attr->recv_cq = qp->ibqp.recv_cq;
1036 	init_attr->srq = qp->ibqp.srq;
1037 	init_attr->xrcd = NULL;
1038 	init_attr->cap = attr->cap;
1039 	init_attr->sq_sig_type = 0;
1040 	init_attr->qp_type = qp->ibqp.qp_type;
1041 	init_attr->create_flags = 0;
1042 	init_attr->port_num = qp->port;
1043 
1044 	mutex_unlock(&qp->mutex);
1045 	return ret;
1046 }
1047